How does Snowflake handle the cost of Anthropic + OpenAI inference at scale?

The Inference Cost Stack Breakdown

The cost of running a single inference call through Snowflake Cortex is composed of several distinct layers, each contributing to the final customer invoice. At the base sits the raw model API cost — the per-token charge from Anthropic or OpenAI. As of Q2 2026, Anthropic's Claude Opus 4 lists at approximately $15 per million input tokens and $75 per million output tokens, while Sonnet 4 runs $3 input and $15 output, and Haiku 4.5 costs $1 input and $5 output. OpenAI's GPT-5 falls in a similar Opus-class band, with o3 priced as a reasoning premium tier and o4-mini serving as the cheap workhorse. This raw token cost is the floor — Snowflake cannot price below it without subsidizing inference.

On top of the model API cost, when Cortex routes Anthropic models via AWS Bedrock, AWS takes a cut on the wholesale rate. This Bedrock passthrough fee is not itemized on Snowflake invoices; it is bundled into the Cortex per-token credit conversion. Industry estimates place this uplift in the low single-digit percentage range based on comparing Bedrock list pricing against Anthropic direct list pricing. Snowflake then adds its own compute fee — the serverless warehouse that runs the Cortex function (CORTEX.COMPLETE, CORTEX.SEARCH, etc.) burns standard credits for orchestration, embedding lookup, and result assembly. While small per call, this cost compounds significantly on agent loops that chain multiple model calls.

The customer-side credit consumption rate determines the final invoice line. Snowflake's credit pricing varies by edition: Standard edition credits run approximately $2, Enterprise edition credits run approximately $3, and Business Critical edition credits run approximately $4. The same query with identical underlying token cost generates different revenue for Snowflake depending on the customer's negotiated credit rate. Finally, governance and observability add micro-credits per call — RBAC checks, prompt logging, PII masking, and Horizon catalog hits all contribute. These are negligible per query but become material at billions of queries per month.

The Margin Math on a 1M-Token Cortex Query

To understand Snowflake's economics, it helps to walk through a concrete example using Claude Opus 4 as the model. All figures here are approximations based on public list pricing as of Q2 2026; actual contract pricing varies materially by customer. Consider a Cortex query that consumes roughly 1 million tokens total. If the full million is split evenly between input and output at the worst case, the raw Anthropic cost hits approximately $90. A more realistic mix of 800,000 input tokens and 200,000 output tokens lands around $27 in raw token cost.

The Bedrock passthrough uplift adds an estimated $0.50 to $1.50 on top of that. Snowflake's serverless compute for orchestration on a single large call runs sub-$1 in credits. On the Enterprise edition at $3 per credit, the customer invoice line for that query lands in the $32 to $38 range based on published Cortex token-to-credit ratios. Snowflake's effective gross margin on the orchestration layer for this call is approximately 10-20% of the line item, with the bulk of the revenue flowing to Anthropic and AWS.

This margin profile is structurally thinner than Snowflake's traditional storage and compute gross margin, which historically runs around 75%. That tension is the central unit-economics challenge for Cortex. The 10-20% margin on partner-model inference is a fraction of what Snowflake earns on its core data warehouse workloads, which means the company must drive high volume to make Cortex a meaningful profit contributor.

Where the Margin Pressure Intensifies

Several specific use cases and customer behaviors create outsized margin pressure on Snowflake's Cortex business. Long-context queries on Opus 4 are particularly challenging — a 200,000-token context window means a single document Q&A call can burn $3 to $5 in raw token cost before any Snowflake margin is applied. Even with the Bedrock uplift and Snowflake's orchestration fee, these calls run on extremely thin margins.

Multi-step agent loops multiply the problem. Cortex Agents that chain 5 to 15 model calls per user request increase the token bill linearly. One "agent question" can equal 20 chat completions in raw cost. RAG over large warehouses introduces another dimension of cost inflation — Cortex Search retrieves chunks and stuffs them into the prompt, and large k-values like top-50 retrieval can inflate input tokens by 10 to 50 times compared to naive prompting.

The customer learning curve creates a variable margin problem. New Cortex customers tend to over-prompt for the first 60 to 90 days — they use no system prompts, no caching, and max-context dumps. During this un-optimized window, Snowflake effectively eats variable margin because the customer's token consumption is higher than necessary. Prompt caching adoption remains a gap — Anthropic's prompt caching can cut repeat-input cost by approximately 90%, but customer adoption inside Cortex lags direct-API users. Until caching is on by default, repeat RAG prompts are paid at full freight.

Reasoning-model premium tiers add another layer of margin pressure. o3 and Opus-extended-thinking calls bill output tokens at the reasoning rate, which can multiply a normal output bill by 3 to 5 times on a single complex query. These calls generate high absolute revenue but thin percentage margins for Snowflake.

The Four Margin-Protection Levers

Snowflake has built four primary mechanisms to protect its margin as inference volume scales. The first and most important is negotiated volume pricing with Anthropic and OpenAI. Snowflake is among the largest enterprise buyers of both Anthropic and OpenAI inference, and enterprise tier discounts off published list pricing are widely assumed in the industry, though not publicly disclosed. The delta between list pricing and Snowflake's wholesale rate is the structural margin floor — the wider that gap, the more room Snowflake has to maintain its 10-20% margin while still offering competitive pricing to customers.

The second lever is the Cortex routing layer, which defaults low-stakes calls to cheaper models before escalating to expensive ones. CORTEX.COMPLETE with model selection, plus newer router functions, automatically send classification, summarization, and extraction tasks to Haiku 4.5, Mistral Large, or Llama 3.x rather than Opus 4. These cheaper models deliver the same answer quality on those tasks at 5 to 15 times lower token cost. The routing is configurable per customer, allowing organizations to define model profiles that map specific use cases to specific price tiers.

The third and most powerful lever is Snowflake Arctic SLM for high-volume low-stakes queries. Arctic and Arctic-Embed run on Snowflake's own infrastructure, meaning there is no per-token payment to a third-party model provider. The margin on Arctic calls approaches Snowflake's traditional compute margin profile of 50-70%. Pushing embedding generation, simple classification, and SQL generation to Arctic is the single biggest margin lever available to Snowflake and its customers.

The fourth lever is customer-side credit guardrails and budget alerts. Resource Monitors on Cortex functions, per-warehouse spend caps, and Snowsight budget alerts do not directly improve Snowflake's margin, but they prevent the customer churn and billing disputes that follow a $400,000 surprise Cortex invoice. By protecting long-term consumption revenue, these guardrails ensure that high-volume customers remain profitable relationships rather than one-time margin events.

What Customers Are Actually Paying in 2026

Public case studies and earnings commentary provide some visibility into real-world Cortex adoption and spending patterns. Bayer has been cited as a public Cortex case study around clinical and research workflows, with Snowflake earnings commentary highlighting life sciences as a leading Cortex vertical, though specific dollar amounts are not disclosed. Siemens is referenced by Snowflake as an enterprise Cortex adopter for industrial and IoT analytics, again without itemized spend.

NYSE and ICE are long-standing Snowflake customers, and Cortex usage patterns in financial services workloads have been emphasized in CFO commentary as a credit-consumption tailwind. Capital One is referenced across Snowflake materials as a flagship financial services account, with Cortex Analyst, Cortex Search, and Cortex Agents named as the expansion vectors, though specific Cortex spend is not public.

The general market signal from Snowflake's Q4 FY26 earnings commentary positioned Cortex as a contributor to product revenue growth without breaking out the specific dollar contribution. The CFO's framing emphasized that Cortex consumption is additive to traditional warehouse spend, not cannibalizing it. This suggests that Cortex is driving net-new workloads rather than simply shifting existing compute budgets.

Cost Optimization via Model Tiering and Caching

Snowflake Cortex implements intelligent model routing that automatically selects the most cost-effective model for each query based on complexity and latency requirements. For high-volume repetitive inference tasks such as classification and summarization, Cortex caches responses at the Snowflake layer, reducing repeated API calls to Anthropic or OpenAI. This caching can cut inference costs by 30-50% for predictable workloads. Customers can also define custom model profiles that map specific use cases to cheaper models — for example, routing internal analytics to Mistral and customer-facing outputs to Claude — further controlling spend without sacrificing quality.

The caching layer is particularly effective for RAG workloads where the same chunks are retrieved repeatedly. When a customer asks similar questions about the same document set, Cortex can serve cached responses rather than re-processing the full context through the model. This reduces both latency and cost, with the savings flowing directly to the customer's bottom line. Snowflake is actively working to make prompt caching the default behavior in Cortex, which would close the adoption gap with direct-API users.

Enterprise Contract Negotiation Leverage

Snowflake's aggregated inference volume across its customer base provides significant negotiating power with model providers. Enterprise customers with multi-year commitments can access custom pricing tiers that undercut public API rates by 20-40% for high-volume usage. Snowflake also offers committed use discounts where customers pre-purchase credits at a 15-25% discount for guaranteed inference volumes. These negotiated rates are transparently passed through to customers via Cortex's billing dashboard, allowing finance teams to track per-model costs and compare against list pricing benchmarks.

The volume aggregation model is a key differentiator for Snowflake versus customers going direct to Anthropic or OpenAI. A single enterprise customer may not have enough volume to negotiate meaningful discounts, but Snowflake's aggregate buying power across thousands of customers creates leverage that benefits the entire customer base. This is particularly valuable for mid-market customers who would otherwise pay full list pricing for API access.

The Arctic SLM Strategy

Snowflake's Arctic SLM represents the company's most strategic margin-protection lever. Arctic and Arctic-Embed run entirely on Snowflake's own infrastructure, meaning there is no per-token payment to a third-party model provider. The margin on Arctic calls approaches Snowflake's traditional compute margin profile of 50-70%, compared to the 10-20% margin on partner-model inference.

The strategy is to push as much high-volume low-stakes work to Arctic as possible. Embedding generation for RAG pipelines, simple classification tasks, and SQL generation are all workloads where Arctic performs comparably to much larger models at a fraction of the cost. Snowflake is investing heavily in Arctic's capabilities to expand the range of tasks it can handle, with the goal of capturing a larger share of inference volume on its own infrastructure.

For customers, the Arctic option provides a clear cost optimization path. Organizations that shift 30-50% of their inference volume from partner models to Arctic can cut their total Cortex spend by 20-40% while maintaining acceptable quality for non-critical workloads. This creates a win-win: customers save money, and Snowflake earns higher margins on the shifted volume.

Related questions

How does Snowflake's Cortex pricing compare to direct API access from Anthropic or OpenAI?

Direct API access typically offers lower per-token costs but lacks Cortex's built-in governance, cost guardrails, and model routing. Snowflake's value is in simplifying access and providing enterprise controls, not in being the cheapest raw token provider.

What is the margin difference between Snowflake's traditional workloads and Cortex inference?

Snowflake's traditional storage and compute gross margin runs approximately 75%, while Cortex partner-model inference margin is approximately 10-20%. Arctic SLM calls approach the traditional margin profile at 50-70%.

Can customers use Snowflake Cortex with models other than Anthropic and OpenAI?

Yes, Cortex supports Mistral, Meta's Llama models, and Snowflake's own Arctic SLM. The routing layer can automatically select the most cost-effective model for each query based on complexity and latency requirements.

How do budget guardrails work in Snowflake Cortex?

Resource Monitors on Cortex functions allow customers to set per-warehouse spend caps and receive Snowsight budget alerts. These can throttle or cap spend in near-real time to prevent runaway costs.

FAQ

Does Snowflake make more margin on Anthropic or OpenAI inference? Snowflake's margin is similar across both providers, typically 10-20% on the inference cost passed through as credits. The margin comes from the orchestration and governance layer, not from marking up the model tokens themselves. Actual margin can vary based on the customer's negotiated credit rate and volume commitments.

Can customers avoid Snowflake's margin by using Anthropic/OpenAI directly? Yes, but they lose Cortex's built-in governance, cost guardrails, and model routing. Snowflake's value is in simplifying access, not in being the cheapest raw token provider. For high-volume low-stakes workloads, the Arctic SLM can be cheaper than any third-party model.

How does Snowflake's Cortex routing layer reduce costs at scale? It automatically defaults expensive calls to cheaper models when latency and quality requirements allow. This can cut per-query costs by 50-90% for many use cases, depending on the workload. The routing is configurable per customer through custom model profiles.

What happens if a customer's inference spend spikes unexpectedly? Snowflake provides budget guardrails that can throttle or cap spend in near-real time. These prevent runaway costs from becoming a margin event for Snowflake or a surprise bill for the customer. Thresholds are set per account or per workload.

Are there volume discounts for large-scale inference customers? Yes, Snowflake negotiates enterprise volume tiers with Anthropic and OpenAI that beat published list pricing. These savings are partially passed through to customers, but Snowflake retains its typical 10-20% margin on the final credit charge. Exact discounts vary by contract.

Does Snowflake charge extra for using Bedrock passthrough versus direct API? Snowflake does not publicly disclose a separate Bedrock passthrough fee. The cost is bundled into the credit charge for the inference call. In practice, the margin remains in the 10-20% range, though Bedrock's own fees may slightly reduce Snowflake's effective margin on those calls.

Sources

• https://www.anthropic.com/pricing
• https://openai.com/pricing
• https://docs.snowflake.com/en/user-guide/snowflake-cortex
• https://aws.amazon.com/bedrock/pricing/
• https://www.snowflake.com/blog/snowflake-cortex-ai/
• https://www.bessemer.com/state-of-the-cloud

Related on PULSE

• [How does Salesforce handle the cost of OpenAI plus Anthropic API spend at scale?](/knowledge/q1557)
• [How do AI inference costs and AI product gross margins work in 2027?](/knowledge/q13059)
• [How do you optimize LLM inference cost in production in 2027?](/knowledge/q12293)
• [Why are 2027 B2B buyers demanding AI-generated demo personalization at scale?](/knowledge/q16516)
• [Why Are GTM Leaders Rethinking Account-Based Strategies as AI Personalizes Outreach at Scale in 2027?](/knowledge/q16247)
• [How Do I Score My Inside and Outside Reps on the Same Scale?](/knowledge/q16051)